Evolutionary search for robust solutions

Information

  • Patent Grant
  • 7783583
  • Patent Number
    7,783,583
  • Date Filed
    Wednesday, September 6, 2006
    18 years ago
  • Date Issued
    Tuesday, August 24, 2010
    14 years ago
Abstract
One embodiment of the present invention is a method for optimizing a parameter set comprising object parameters, the method comprising the steps of: (a) creating an initial population of a plurality of individual parameter sets, the parameter sets comprising object parameters describing a model, structure, shape, design or process to be optimized and setting the initial population as a current parent population; (b) for each individual parameter set in a parent population mutating the parameters and optionally recombining the parameters to create an offspring population of individual parameter sets, wherein the strength of an individual object parameter mutation is enlarged by a noise contribution to enhance the robustness of the optimization; (c) evaluating a quality of each individual in the offspring population; (d) selecting individuals of the offspring population to be the next parent generation; and (e) repeating steps (b) through (d) until a termination criterion is reached.
Description
RELATED APPLICATIONS

This application is related to U.S. patent application Ser. No. 10/007,906 and U.S. patent application Ser. No. 10/080,742 which are incorporated by reference herein in their entirety.


FIELD OF THE INVENTION

The present invention relates generally to the optimization of designs, models, structures and. shapes, for example in fluid-dynamic applications such as the design of an aircraft wing, a gas turbine or a compressor blade. In particular, it relates to the optimization of such designs using Evolutionary Algorithms (EAs).


BACKGROUND OF THE INVENTION

In the field of evolutionary algorithms, basic principles of natural evolution are used for generating or optimizing technical structures. Basic operations are mutation and recombination as a method for modifying structures or parameters. To eliminate unfavorable modifications and to proceed with modifications which increase the overall quality of the system, a selection operation is used. Principles of the evolution strategy can be found for example in Rechenberg, Ingo (1994) “Evolutionsstrategie”, Friedrich Frommann Holzboog Verlag, which is incorporated by reference herein in its entirety.


The application of EAs in the optimization of designs is well known, see for example the book “Evolutionary Algorithms in Engineering Applications” by D. Dasgupta and Z. Michalewicz, Springer Verlag, 1997, which is incorporated by reference herein in its entirety.


As evolutionary algorithms are more and more successfully used as optimization tools for large-scale “real-world” problems, the influence of noise on the performance and the convergence properties of evolutionary algorithms have come into focus.


Quality evaluations in optimization processes are frequently noisy due to design uncertainties concerning production tolerances or actuator imprecision acting directly on the design variables x. That is, the performance ƒ of a design becomes a stochastic quantity {tilde over (ƒ)} via internal design perturbations

{tilde over (ƒ)}(x)=ƒ(x+δ), δ-random vector,

where the random vector δ obeys a certain unknown distribution (often modeled as a Gaussian distribution) and

E[δ]=0.


This means, given a design x, evaluating its quality {tilde over (ƒ)}(x) necessarily yields stochastic quantity values. As a result, an optimization algorithm applied to {tilde over (ƒ)}(x) must deal with these uncertainties and it must use this information to calculate a robust optimum based on an appropriate robustness measure.


Probably the most widely used measure is the expected value of {tilde over (ƒ)}(x), that is







E


[


f
~



|


x

]


.





Assuming a continuous design space, the expected value robustness measure is given by the integral







E


[


f
~



|


x

]


=





N





f


(

x
+
δ

)




p


(
δ
)










N


δ








and the optimal design x is formally obtained by







x
^

=

arg






opt
x







N





f


(

x
+
δ

)




p


(
δ
)











N


δ

.








If one were able to calculate






E


[


f
~



|


x

]






analytically, the resulting optimization problem would be an ordinary one, and standard (numerical) optimization techniques could be applied. However, real-world applications will usually not allow for an analytical treatment, therefore one has to rely on numerical estimates of






E


[


f
~



|


x

]






using Monte-Carlo simulations. Alternatively one can use direct search strategies capable of dealing with the noisy information directly.


The latter is the domain of evolutionary algorithms (EAs). In particular, evolutionary algorithms have been shown to cope with such stochastic variations better than other optimization algorithms, see e.g. “On the robustness of population-based versus point-based optimization in the presence of noise” by V. Nissen and J. Propach, IEEE Transactions on Evolutionary Computation 2(3):107-119, 1998, which is incorporated by reference herein in its entirety.


A conventional technique to find approximate solutions to the above equation using EAs is to use the design uncertainties δ explicitly. That is, given an individual design x, the perturbation δ is explicitly added to the design x. While the EA works on the evolution of x, the goal function in the black-box is evaluated with respect to {tilde over (x)}:=x+δ. Since in center of mass evolution strategies an individual offspring design is the result of a mutation z applied to a parental individual and the parental centroid custom characterxcustom character), respectively, the actually design tested is

{tilde over (x)}:=custom characterxcustom character+z+δ.


Taking now another perspective, one might interpret z+δ as a mutation in its own right. This raises the question whether it is really necessary to artificially add the perturbation in a black-box to the design x. As an alternative one might simply use a mutation {tilde over (z)}=z+δ with a larger mutation strength instead of z. In other words, the mutations themselves may serve as robustness tester.


However, even though evolutionary algorithms/strategies are regarded as well suited for noisy optimization, its application to robust optimization bears some subtleties/problems: due to selection, the robustness of a design x is not tested with respect to samples of the density function p(δ). Selection prefers those designs which are by chance well adopted to the individual realizations of the perturbation δ.


For example, when considering actuator noise of standard deviation ε on a sphere model ∥x∥2 (to be minimized), the actually measured standard deviation Di of a specific component i of the parent population will usually be smaller, i.e. Di<σ. This is so, because selection singles out all those x+δ states with large length ∥x+δ∥. That is, shorter δ vectors are preferred resulting in a smaller measured standard deviation. Therefore, an evolutionary algorithm for robust optimization must take into account this effect and take appropriate counter measures.


Furthermore, it is well known that noise deteriorates the performance of the evolutionary algorithms. If the function to be optimized is noisy at its global or local optimizer, the evolutionary algorithm cannot reach the optimizer in expectation. That is, the parental individuals are located in the long run (steady state behavior) in a certain (expected) distance to the optimizer, both in the object parameter space and usually also in the quality/fitness space.


What is needed is an improved system and method (1) for evaluating the robustness of an Evolutionary Algorithm; (2) where the observed parental variance is controlled such that robustness (with regard to noise etc.) is tested correctly; and/or (3) for optimization that is driven by the trade-off between reducing the residual distance (induced by the noise) to the optimizer state and reducing the number of required fitness evaluations. In other words, such a method for optimization can reduce the residual distance (induced by the noise) to the optimizer state while at the same time minimizing the required additional fitness evaluation effort.


SUMMARY OF THE INVENTION

A system and method (1) for evaluating the robustness of an Evolutionary Algorithm; (2) where the observed parental variance is controlled such that robustness (with regard to noise etc.) is tested correctly; and/or (3) for optimization that is driven by the trade-off between reducing the residual distance (induced by the noise) to the optimizer state and reducing the number of required fitness evaluations. In other words, such a method for optimization can reduce the residual distance (induced by the noise) to the optimizer state while at the same time minimizing the required additional fitness evaluation effort.


One embodiment of the present invention is a method for optimizing a parameter set comprising object parameters, the method comprising the steps of: (a) creating an initial population of a plurality of individual parameter sets, the parameter sets comprising object parameters describing a model, structure, shape, design or process to be optimized and setting the initial population as a current parent population; (b) for each individual parameter set in a parent population mutating the parameters and optionally recombining the parameters to create an offspring population of individual parameter sets, wherein the strength of an individual object parameter mutation is enlarged by a noise contribution to enhance the robustness of the optimization; (c) evaluating a quality of each individual in the offspring population; (d) selecting individuals of the offspring population to be the next parent generation; and (e) repeating steps (b) through (d) until a termination criterion is reached.


The features and advantages described in the specification are not all inclusive and, in particular, many additional features and advantages will be apparent to one of ordinary skill in the art in view of the drawings, specification, and claims. Moreover, it should be noted that the language used in the specification has been principally selected for readability and instructional purposes, and may not have been selected to delineate or circumscribe the inventive subject matter.





BRIEF DESCRIPTION OF THE FIGURES


FIG. 1 is a schematic flow diagram of the optimization method according one embodiment of the present invention



FIG. 2 is a schematic flow diagram of the optimization method according to another embodiment of the present invention.



FIG. 3 is pseudocode representing one implementation according to one embodiment of the present invention.



FIG. 4 is pseudocode representing one implementation according to another embodiment of the present invention.





DETAILED DESCRIPTION OF THE INVENTION

A preferred embodiment of the present invention is now described with reference to the figures where like reference numbers indicate identical or functionally similar elements. Also in the figures, the left most digit of each reference number corresponds to the figure in which the reference number is first used.


Reference in the specification to “one embodiment” or to “an embodiment” means that a particular feature, structure, or characteristic described in connection with the embodiments is included in at least one embodiment of the invention. The appearances of the phrase “in one embodiment” in various places in the specification are not necessarily all referring to the same embodiment.


Some portions of the detailed description that follows are presented in terms of algorithms and symbolic representations of operations on data bits within a computer memory. These algorithmic descriptions and representations are the means used by those skilled in the data processing arts to most effectively convey the substance of their work to others skilled in the art. An algorithm is here, and generally, conceived to be a self-consistent sequence of steps (instructions) leading to a desired result. The steps are those requiring physical manipulations of physical quantities. Usually, though not necessarily, these quantities take the form of electrical, magnetic or optical signals capable of being stored, transferred, combined, compared and otherwise manipulated. It is convenient at times, principally for reasons of common usage, to refer to these signals as bits, values, elements, symbols, characters, terms, numbers, or the like. Furthermore, it is also convenient at times, to refer to certain arrangements of steps requiring physical manipulations of physical quantities as modules or code devices, without loss of generality.


However, all of these and similar terms are to be associated with the appropriate physical quantities and are merely convenient labels applied to these quantities. Unless specifically stated otherwise as apparent from the following discussion, it is appreciated that throughout the description, discussions utilizing terms such as “processing” or “computing” or “calculating” or “determining” or “displaying” or “determining” or the like, refer to the action and processes of a computer system, or similar electronic computing device, that manipulates and transforms data represented as physical (electronic) quantities within the computer system memories or registers or other such information storage, transmission or display devices.


Certain aspects of the present invention include process steps and instructions described herein in the form of an algorithm. It should be noted that the process steps and instructions of the present invention could be embodied in software, firmware or hardware, and when embodied in software, could be downloaded to reside on and be operated from different platforms used by a variety of operating systems.


The present invention also relates to an apparatus for performing the operations herein. This apparatus may be specially constructed for the required purposes, or it may comprise a general-purpose computer selectively activated or reconfigured by a computer program stored in the computer. Such a computer program may be stored in a computer readable storage medium, such as, but is not limited to, any type of disk including floppy disks, optical disks, CD-ROMs, magnetic-optical disks, read-only memories (ROMs), random access memories (RAMs), EPROMs, EEPROMs, magnetic or optical cards, application specific integrated circuits (ASICs), or any type of media suitable for storing electronic instructions, and each coupled to a computer system bus. Furthermore, the computers referred to in the specification may include a single processor or may be architectures employing multiple processor designs for increased computing capability.


The algorithms and displays presented herein are not inherently related to any particular computer or other apparatus. Various general-purpose systems may also be used with programs in accordance with the teachings herein, or it may prove convenient to construct more specialized apparatus to perform the required method steps. The required structure for a variety of these systems will appear from the description below. In addition, the present invention is not described with reference to any particular programming language. It will be appreciated that a variety of programming languages may be used to implement the teachings of the present invention as described herein, and any references below to specific languages are provided for disclosure of enablement and best mode of the present invention.


In addition, the language used in the specification has been principally selected for readability and instructional purposes, and may not have been selected to delineate or circumscribe the inventive subject matter. Accordingly, the disclosure of the present invention is intended to be illustrative, but not limiting, of the scope of the invention, which is set forth in the claims.


1. Notational Conventions


In order to unify/simplify the notations for the pseudo code description of the algorithms the following conventions will be used: g is the generation (time) counter, it appears as parenthesized superscript on the respective quantities. The i-th component of a vector x is denoted by xi. Vectors are always denoted by boldface letters in the following. In order to refer to the i-th component of a vector x we will alternatively write (x)i, that is (x)i≡xi. N denotes the object parameter space dimension. xεR N is the N-dimensional object parameter vector. σ is referred to as the mutation strength being the standard deviation of the normally distributed mutations. μ is the parental population size. Quantities related to parental individuals are indexed by subscript m. λ is the number of offspring generated in a single generation. Quantities related to offspring individuals are indexed by subscript l and are denoted with a tilde. θ is an exogenous strategy parameter called truncation ratio, defined as






ϑ
:=

μ
λ





Normally distributed random variables/numbers y are denoted by N( y2) where y is the expected value of y and σ its standard deviation. A vector y of normally distributed random variables is symbolized by N( y; V), where y is the expected value vector of y and V stands either for the covariance matrix C or the vector of standard deviations, whereby the off-diagonal elements of C are assumed to be zero and the diagonal elements are ε2i. In the latter case the density function reads for the actuator noise strength vector ε. The subscript notation m; λ denotes quantities of the mth-best individual, i.e., that individual being the mth-largest (in the case of maximization) or smallest (in the case of minimization) with respect to its observed (measured) fitness ƒ(x). custom characterycustom character) denotes the parental population average, i.e.








y


:=


1
μ






m
=
1

μ




y
m

.







This is basically a centroid calculation. Overlined symbols, such as xi, when used in an algorithm, are used to denote averaging over time, i.e. usually this is a weighted average over the generations g. R is the length of the centroid state

R:=∥custom characterxcustom character∥.


r is the length of the first N−1 components of the vector x






r
:=





i
=
1


N
-
1




x
i
2








FIG. 1 shows a first embodiment of the invention, which takes advantage of the idea of direct robustness testing via mutations.


a. Parameter Iinitialization


In step 100, the basic parameters of the optimization algorithm are initialized.


The generation g is set to a start value (0). The mutation strength σ, the population size λ and the number of offspring generated in a single generation λ are all set to initial values, whereby the number of offspring depends on the truncation ratio θ. The recombinant custom characterxcustom character is set to an initial vector x(init).


b. Procreation of Offspring for Generation n


In step 110, the procreation of the λ offspring is realized.


First, a log-normal mutation of the recombined strategy parameter σ is performed:

σl:=custom characterσcustom characterexp[τσNl(0,1)]  (1)

where τσ is an exogenous strategy parameter, the so-called learning parameter. In order to ensure linear convergence order of the evolutionary algorithm on the sphere, it is known to be sufficient to ensure that τσ∝1/√{square root over (N)}, such that










τ
σ

=

1

N






(
2
)








is a reasonable choice.


The new recombination of the strategy parameter is calculated as











σ


:=


1
μ






m
=
1

μ




σ
~


m
;
λ








(
3
)







Secondly, a mutation of the object parameter is performed on top of the recombinant custom characterxcustom character:

({tilde over (x)}l)i:=custom characterxicustom character+√{square root over ({tilde over (σ)}l2l2)}Nl,i(0,1)  (4)

where l=1, . . . , λ and i=1, . . . , N and the new recombinant custom characterxcustom character is calculated as











x


:=


1
μ






m
=
1

μ




x
~


m
;
λ








(
5
)







At this point, it is to be noted that the actual strength by which the mutation is performed differs from the known “standard” evolution strategy: In order to account for noise (actuator noise etc.), the strength consists of the evolution strategy specific contribution σ and an additional noise contribution ε. Since normality of the noise is assumed, the sum of the strategy-specific mutation contribution and the noise contribution is still a normally distributed random vector, however, with variance {tilde over (σ)}l2i2for the i-th component. Performing the mutations in this way allows for taking advantage of the evolution strategy immanent mutation (of strength σl) as an additional robustness tester. That is, there is no need to evolve the EA's mutation strength down to very small values since it can take over a part of the robustness testing itself.


c. Controlling Mutation Strength (Robustness Variance Control)


In step 120, the strength of the mutation of the object parameter (cf. equation 4) is controlled by (a) measuring the parental population variance or its standard deviation Di and (b) adjusting the εi accordingly.


With regard to substep (a), it must be noted that, since robustness testing is highly noisy, calculating the parental population variance from just one generation results in highly fluctuating Di estimates not well suited for εi control. Therefore, a smooth Di estimate is needed. One way of smoothing the data is by weighted accumulation, also known as exponential averaging. Since

Di=√{square root over ( xi2xi2)}  (6)

Di can be obtained from the smoothed time averages of xi and xi2. The exponential averaging is designed in such a way that the xi and xi2 information fades away exponentially fast if custom characterxicustom character and custom characterxi2custom character, respectively, are zero. The time constant by which this process happens is controlled by the accumulation time constant cxε[0,1]. Since the changing rates of the evolution strategy (e.g. the progress rate on the sphere) are often of the order 1/N, it is reasonable to use










c
x

=

1
N





(
7
)








as a first choice.


With regard to substep (b), it is to be noted that the actual mutation strength depends on σ and εi. It is important to realize that εi is not equivalent to the desired actuator noise strength εi*. The latter is the desired strength by which the actually realized design instances are tested. As was already stated in the beginning, due to the (μ,λ)-selection, the actual variances of the selected (i.e. parental) {tilde over (x)}m;λ states are usually smaller than the desired ε1*.


Therefore, εi must be controlled in such a way that the observed (i.e. measured) standard deviation

Di:=√{square root over (Var[{({tilde over (x)}1;λ)i, . . . , ({tilde over (x)}μ;λ)i}])}  (8)

gets close to εi.


Given a stable estimate of the real parental population variance, one can compare it with the desired noise strength εi*.


That is, the aim is to control the observed parental population variance Di in such a way that

Di≈εi*  (9).


If the evolution strategy is able to get close to the robust optimizer, then the above condition ensures that robustness is guaranteed for the correct target noise strength. While, in general, one cannot be sure that the evolution strategy locates the robust optimizer, fulfilling the above condition can be ensured by the control rule

εi:=εi exp└τεSign(εi*−Di)┘  (10).


If Dii*, the above equation does not change εi. In case Dii*, εi is increased and if Dii*, εi is decreased. Due to the choice of the sign function, the ε change rate is independent of the actual value of the Di−εi* difference. This ensures that large differences do not result in extreme εi changes. As an alternative one might replace sign( ) by a sigmoid function, e.g. the hyperbolic tangent.


The choice of the parameter τε, which may be interpreted as a damping constant, must be taken with care. The dynamics of the Di and εi interfere with each other. As a result, such a system can exhibit instabilities, e.g. oscillatory behavior. In order to prevent such instabilities, cx and τε must be chosen appropriately. While there is clearly a need for a thorough analysis, in the investigations done so far, the choice










τ
ɛ

=


1
3



c
x






(
11
)








worked flawlessly.


d. Increasing/Controlling Population Size


In step 130 as shown in FIG. 1, it is checked whether the population size is to be adapted or not.


In order to control the population size λ, a measure is needed which allows for a decision whether to increase λ. Assuming a stationary actuator noise distribution, the dynamics of the evolution strategy will (usually) approach a steady state behavior in a certain vicinity of the optimizer. That is, for a certain time period one observes on average a measurable improvement in the observed parental fitness values. If, however, one reaches the vicinity of the steady state, parental fitness will start to fluctuate around an average value. Therefore, if one observes on average no improvements of the fitness values from generation g to g+1, it is time to increase the population size (a rule for λ-decrease has not been developed so far). The average parental fitness change ΔF is given by

ΔF=custom characterFcustom character(g)custom characterFcustom character(g−1)  (12)

where











F


:=


1
μ






m
=
1

μ





F
~


m
;
λ


.







(
13
)







Since ΔF itself is a strongly fluctuating quantity, an exponential smoothing should be used to avoid unnecessary population increase due to random fluctuations, as in line 28 of the algorithm shown in FIG. 3, where cf determines the time constant by which ancient ΔF-information vanishes exponentially fast. As a natural choice,










c
f

:=

1
N





(
14
)








can be used.


As an update rule, the population size is increased at every Δg-th generation, if the (exponentially smoothed) average fitness change ΔF is smaller than 0, whereby Δg is an exogenous strategy parameter.


When considering maximization, desired fitness changes are of the kind ΔF>0. Therefore, if ΔF≦0, the population size λ should be increased. Since the increased population does not necessarily change the sign of ΔF in the next generation (random fluctuations), the test of the update rule in Line 29 is performed every Δg-th generation.


The λ-update itself is carried out in step 140 of FIG. 1 using the following formula:

μ:=|μcμ|  (15).


That is, the new μ is obtained from the old μ using the change rate cμ.


e. Terminating Condition


In step 150 it is checked whether the optimized state fulfils a predetermined criterion.


If yes, the procedure ends. If no, the procedure continues with the procreation of offspring in the next generation.


As termination conditions the standard stopping rules can be used:

    • Resource criteria:
      • Maximum number of generations
      • Maximum CPU-time
    • Convergence criteria:
      • In the space of fitness values,
      • In the space of object parameters,
      • In the space of strategy parameters.


        f. Concrete Implementation


A concrete example of an implementation of the above-described method is presented using pseudocode in FIG. 3 of this description, which is hereby incorporated into this description by reference.


Parameter initialization (step 100 in FIG. 1) is done on lines 1 to 10.


The procreation of offspring (step 110 in FIG. 1) is done on lines 11 to 16, with line 13 performing the log-normal mutation of recombined strategy parameter σ. In line 14, the mutation of the object parameters is performed according to the invention.


An intermediate recombination of the strategy parameter σ is done on line 18 of FIG. 3, since it needs to be calculated only once per generation. The intermediate recombinant custom characterxcustom character is calculated on line 19 of FIG. 3.


Control of the actually observed fluctuation strength (step 120 in FIG. 1) is realized according to the invention on lines 20 to 26 of FIG. 3.


Control of the population size (steps 130 and 140) is realized according to the invention in lines 27 to 32 of FIG. 3.


There are three new exogenous strategy parameters to be fixed: The truncation ratio θ:=μ/λ, the update time interval Δg, and the μ change rate cμ. Extensive simulations suggest

Δg=N, cμ=4

as a rule of thumb. That is, population upgrading should be done in a rather aggressive manner. The choice of the truncation ratio should be in the interval

θ=0.4, . . . , 0.6.


The second embodiment of the invention shown in FIG. 2 is an adaptation of a different evolutionary strategy, known in the literature as cumulative step size adaptation (CSA).


Unlike the mutation strength adaptation by the previous embodiment described with reference to FIG. 1 that uses one-generation fitness ranking information only, the cumulative step size adaptation strategy relies on quality (fitness) related search space information gathered over a sequence of consecutive generations. In these strategies, the length of so-called evolution paths is used to control the variance σ2 of the object parameter mutation operator. For a comprehensive introduction to CSA see Hansen, N. and Ostermeier, A, “Completely Derandomized Self-Adaptation in Evolution Strategies”, in: Evolutionary Computation, 9(2), p. 159 to 195.


Due to the special way how the mutation strength is determined in the CSA evolution strategy and how the offspring are generated using the same mutation strength σ for all offspring individuals, there seems to be no direct way to transfer the idea of direct robustness testing through mutations to the CSA evolution strategy.


Therefore, the present embodiment employs a black-box approach: the evolution strategy is applied without modifications to the function ƒ(x) which is internally disturbed by actuator noise of strength ε. Thus, one has to differentiate between the evolution strategies' individual vectors {tilde over (x)}l and the real actuator state {tilde over (x)}la entering the ƒ function. The latter is invisible to the employed evolution strategy, however, it is needed for calculating the actually realized parental actuator fluctuations measured by the standard deviation Di.


In order to perform the path accumulation, the accumulation time constant cσ must be fixed. Two different recommendations concerning cσ can be found in the literature: ∝1/√{square root over (N)} and ∝1/N. From the viewpoint of stability







c
σ

=

1
N






should be chosen. According to experimental evidences and theoretical analysis, the damping constant dσ must be chosen depending on cσ







d
σ

=

1

c
σ






A concrete example of an implementation of the method described in relation to FIG. 2 is presented using pseudo-code in FIG. 4 of this description, which is hereby incorporated into this description by reference.


Parameter initialization (step 200 in FIG. 2) is done on lines 1 to 11.


The procreation of offspring (step 210 in FIG. 2) is done on lines 12 to 17. The evolution strategies' individual vectors {tilde over (x)}l are generated in Line 14 of FIG. 4 and the real actuator state {tilde over (x)}la entering the ƒ function is generated on Line 15 of FIG. 4.


The major difference to the method according to the first embodiment is located in Lines 20 and 21 of FIG. 4: the path cumulation and the a update.


Control of the actually observed fluctuation strength (step 220 in FIG. 1) is realized according to the invention on lines 22 to 29 of FIG. 4.


Control of the population size (steps 230 and 240) is realized according to the invention in lines 32 to 35 of FIG. 4.


The rest of the algorithm in FIG. 4 is directly taken from the first embodiment shown in FIG. 1 and FIG. 3. The same holds for the recommended choice of the endogenous strategy parameters cx, cf, cμ, and θ.


The present invention finds application for all kinds of structure encoding.


Specific examples of application of the present invention apart from turbine blades are airfoil shapes and other aerodynamic or hydrodynamic structures. Other fields of application are architecture and civil engineering, computer science, wing design, engine valve design and scheduling problems.


While particular embodiments and applications of the present invention have been illustrated and described herein, it is to be understood that the invention is not limited to the precise construction and components disclosed herein and that various modifications, changes, and variations may be made in the arrangement, operation, and details of the methods and apparatuses of the present invention without departing from the spirit and scope of the invention as it is defined in the appended claims.

Claims
  • 1. A computer based method for optimizing a design of a hydrodynamic or aerodynamic structure, the method comprising the steps of: (a) creating an initial population of a plurality of individual parameter sets the parameter sets comprising object parameters describing the design to be optimized and setting the initial population as a current parent population;(b) for each individual parameter set in a parent population mutating, by the computer, the parameters and optionally recombining the parameters to create an offspring population of individual parameter sets, wherein a strength of an individual object parameter mutation is enlarged by a noise contribution, wherein the noise contribution is added to said individual object parameter value whereby the noise contribution is a mutation term, to enhance the robustness of the optimization;(c) evaluating, by the computer, a quality of each individual in the offspring population;(d) selecting individuals of the offspring population to be the next parent generation based upon said quality; and(e) repeating steps (b) through (d) until a termination criterion is reached to optimize the design.
  • 2. The method of claim 1 wherein said parameter set includes strategy parameters.
  • 3. The method of claim 1, wherein the strength of the noise contribution is adapted such that the estimated population variance substantially equals a prescribed variance governed by a robustness criterion.
  • 4. The method of claim 1, wherein the noise contribution varies for the different object parameters of a given parameter set.
  • 5. The method of claim 1 wherein the selection in step (c) is a deterministic (μ,λ) or (μ+λ) selection of evolution strategies.
  • 6. The method of claim 1 wherein the center of mass recombination known in evolution strategies as (μ/μ,λ) is used in step (b).
  • 7. The method of claim 1 wherein the random source for mutation is normally distributed.
  • 8. The method of claim 1, wherein the estimation of the population variance is subject to exponential smoothing.
  • 9. A computer including a software program, which when executed by a computing device performs a method according to claim 1.
  • 10. An apparatus programmed for executing a method according to claim 1.
  • 11. A computer for optimizing a design of a hydrodynamic or aerodynamic structure, comprising: parameter sets means for creating an initial population of a plurality of individual parameter sets,the parameter sets comprising object parameters describing the design to be optimized and possibly strategy parameters and setting the initial population as the current parent population;mutating means, for each individual parameter set in the parent population, mutating the parameters and optionally recombining them, to create an offspring population of individual parameter sets, wherein a strength of an individual object parameter mutation is enlarged by a noise contribution, wherein the noise contribution is added to said individual object parameter value whereby the noise contribution is a mutation term, to enhance the robustness of the optimization;evaluating means for evaluating a quality of each individual in the offspring population;selecting means for selecting individuals of the offspring population to be the next parent generation based upon said quality; andterminating means for repeatedly implementing the mutating means, evaluating means and selecting means until a termination criterion is reached for optimizing the design;wherein the size of the population of parameter sets is varied, preferably increased depending on an average change in the evaluated quality with respect to a preceding population of parameter sets.
  • 12. The system of claim 11, wherein said parameter set includes strategy parameters.
  • 13. The system of claim 11, wherein the average change in quality is subject to exponential smoothing before deciding on a variation of the population size.
  • 14. The system of claim 11, wherein the selection in step (c) is a deterministic (μ,λ) or (μ+λ) selection of evolution strategies.
  • 15. The system of claim 11, wherein the center of mass recombination known in evolution strategies as (μ/μ,λ) is used in step (b).
Priority Claims (1)
Number Date Country Kind
05019800 Sep 2005 EP regional
US Referenced Citations (43)
Number Name Date Kind
5074752 Murphy et al. Dec 1991 A
5136686 Koza Aug 1992 A
5148513 Koza et al. Sep 1992 A
5265830 Allen Nov 1993 A
5319781 Syswerda Jun 1994 A
5355528 Roska et al. Oct 1994 A
5461570 Wang et al. Oct 1995 A
5487130 Ichimori et al. Jan 1996 A
5541848 McCormack et al. Jul 1996 A
5724258 Roffman Mar 1998 A
5819244 Smith Oct 1998 A
5924048 McCormack et al. Jul 1999 A
6285968 Motoyama et al. Sep 2001 B1
6292763 Dunbar et al. Sep 2001 B1
6430993 Seta Aug 2002 B1
6449603 Hunter Sep 2002 B1
6516309 Eberhart et al. Feb 2003 B1
6549233 Martin Apr 2003 B1
6578018 Ulyanov Jun 2003 B1
6606612 Rai et al. Aug 2003 B1
6654710 Keller Nov 2003 B1
6662167 Xiao Dec 2003 B1
6781682 Kasai et al. Aug 2004 B1
6879388 Kasai et al. Apr 2005 B2
6928434 Choi et al. Aug 2005 B1
6950712 Ulyanov et al. Sep 2005 B2
7043462 Jin et al. May 2006 B2
7047169 Pelikan et al. May 2006 B2
20020099929 Jin et al. Jul 2002 A1
20020138457 Jin et al. Sep 2002 A1
20020165703 Olhofer et al. Nov 2002 A1
20030030637 Grinstein et al. Feb 2003 A1
20030055614 Pelikan et al. Mar 2003 A1
20030065632 Hubey Apr 2003 A1
20030191728 Kulkami et al. Oct 2003 A1
20040014041 Allan Jan 2004 A1
20040030666 Marra et al. Feb 2004 A1
20040034610 De Lacharriere et al. Feb 2004 A1
20040049472 Hayashi et al. Mar 2004 A1
20040138846 Buxton et al. Jul 2004 A1
20050197994 Fujii et al. Sep 2005 A1
20050209982 Jin et al. Sep 2005 A1
20050246297 Chen et al. Nov 2005 A1
Foreign Referenced Citations (3)
Number Date Country
1205863 May 2002 EP
1205877 May 2002 EP
WO 02057946 Jul 2002 WO
Related Publications (1)
Number Date Country
20070094167 A1 Apr 2007 US